No Filter: Cultural and Socioeconomic Diversityin Contrastive Vision-Language Models

Cultural and Socioeconomic Diversity in Vision-Language Models: A Look Under the Hood

Introduction

Vision-Language Models (VLMs) like CLIP and SigLIP have become indispensable in bridging the gap between visual content and textual information. While these models have achieved impressive performance on standard benchmarks, their cultural and socioeconomic biases often go unnoticed. This paper explore these biases, examining how training on English-only datasets impacts cultural understanding and proposing ways to improve VLMs' global representation.

Issues with English-only Training Data

A Case Against English-only Data

One major finding is that training VLMs solely on English image-text pairs can disadvantage communities with lower socioeconomic status and reduce the models' ability to understand cultural contexts. This becomes evident when these models are evaluated using non-Western benchmarks like the Google Landmarks Dataset (GLDv2) and Dollar Street, which include images and text originating from various socioeconomic and geographical backgrounds.

Highlighting the Bias

For example, VLMs trained on English-only data often misidentify landmarks from non-Western countries as similar-looking landmarks in English-speaking nations. When tested on the Dollar Street dataset, which spans a wide range of household items from different income levels, models pretrained on global data performed substantially better than those trained only on English data, particularly for images from lower-income households.

Bridging the Gap with Diverse Training Data

Pretraining on Global Data

The research found that using globally diverse training data before fine-tuning on English content can improve models' cultural understanding without losing performance on popular benchmarks like ImageNet and COCO. In other words, a model trained initially on a mixture of languages and then fine-tuned on English data can offer the best of both worlds: higher cultural diversity without sacrificing accuracy on Western-oriented benchmarks.

Figures of Improvement

• Dollar Street Zero-Shot Accuracy: Training on global data improved performance from 48.52% to 49.96%.
• GLDv2 Zero-Shot Accuracy: Improved from 43.84% to 49.46% when global data was used.
• ImageNet Zero-Shot: While there was a slight drop in accuracy (from 70.36% to 68.23%), the benefits in terms of cultural representation were substantial.

Introducing Geo-localization as a Metric

The paper introduces geo-localization as a novel metric for assessing VLMs' cultural diversity. This task involves predicting the geographical origin of an image based on its visual features. The findings suggest that models trained on English-only data struggle significantly in this task compared to those trained on global data.

Few-shot Geo-localization

Using a linear probe on the image encoder, it was found that the geo-localization accuracy for models trained on global data was significantly higher. For instance, when comparing models trained on English-only data versus those trained on globally diverse data, the latter showed improvements of up to 13.30% in few-shot country predictions on the GeoDE dataset.

Decoupling Multilinguality and Cultural Diversity

Despite being useful for evaluating multilingual capabilities, datasets designed to test multilinguality, like XM3600, may not be sufficient for evaluating cultural diversity. The study showed that retrieval tasks using XM3600 did not highlight any significant differences between models trained on different datasets, suggesting that merely translating captions does not adequately capture the cultural nuances that diversified data can offer.

Fine-tuning and Data Mixing

To balance the trade-off between cultural diversity and performance on standard benchmarks, two strategies were proposed:

1. Fine-tuning: Pretraining on globally diverse data and fine-tuning on English-only data can yield a balanced model. Fine-tuning for as few as 50k steps was often sufficient.
2. Data Mixing: Mixing different proportions of data during pretraining can also achieve a good balance, although it requires training new models and is computationally more expensive.

Implications and Future Directions

Practical Implications

Improving the cultural diversity of VLMs has obvious practical benefits. It can lead to more inclusive AI systems that better serve global communities, understand a wider range of socioeconomic contexts, and thus offer more accurate and equitable outcomes.

Theoretical Implications

From a theoretical standpoint, these findings emphasize the importance of diverse training data for building robust and inclusive AI systems. They also introduce geo-localization as a valuable new metric for assessing cultural diversity in VLMs.

Conclusion

This research highlights the importance of using culturally and socioeconomically diverse training data to build more inclusive VLMs. While there is a trade-off between cultural diversity and performance on well-known benchmarks, strategic pretraining and fine-tuning can help achieve a balance. Moving forward, AI practitioners should consider these factors to develop more globally representative models.

In summary, the call to move away from English-only training data in favor of a more globally diverse approach offers the potential for AI systems to better understand and serve our complex, multicultural world.